===================
Features
--------
- Waiters on Event and Semaphore objects that call ``wait()`` or
``acquire()``, respectively, that find the Event already set, or the
Semaphore available, no longer "cut in line" and run before any
previously scheduled greenlets. They now run in the order in which
they arrived, just as waiters that had to block in those methods do.
See :issue:`1520`.
- Update tested PyPy version from 7.3.0 to 7.3.1 on Linux.
See :issue:`1569`.
- Make ``zope.interface``, ``zope.event`` and (by extension)
``setuptools`` required dependencies. The ``events`` install extra now
does nothing and will be removed in 2021.
See :issue:`1619`.
- Update bundled libuv from 1.36.0 to 1.38.0.
See :issue:`1621`.
- Update bundled c-ares from 1.16.0 to 1.16.1.
On macOS, stop trying to adjust c-ares headers to make them
universal.
See :issue:`1624`.
Bugfixes
--------
- Make gevent locks that are monkey-patched usually work across native
threads as well as across greenlets within a single thread. Locks that
are only used in a single thread do not take a performance hit. While
cross-thread locking is relatively expensive, and not a recommended
programming pattern, it can happen unwittingly, for example when
using the threadpool and ``logging``.
Before, cross-thread lock uses might succeed, or, if the lock was
contended, raise ``greenlet.error``. Now, in the contended case, if
the lock has been acquired by the main thread at least once, it should
correctly block in any thread, cooperating with the event loop of both
threads. In certain (hopefully rare) cases, it might be possible for
contended case to raise ``LoopExit`` when previously it would have
raised ``greenlet.error``; if these cases are a practical concern,
please open an issue.
Also, the underlying Semaphore always behaves in an atomic fashion (as
if the GIL was not released) when PURE_PYTHON is set. Previously, it
only correctly did so on PyPy.
See :issue:`1437`.
- Rename gevent's C accelerator extension modules using a prefix to
avoid clashing with other C extensions.
See :issue:`1480`.
- Using ``gevent.wait`` on an ``Event`` more than once, when that Event
is already set, could previously raise an AssertionError.
As part of this, exceptions raised in the main greenlet will now
include a more complete traceback from the failing greenlet.
See :issue:`1540`.
- Avoid closing the same Python libuv watcher IO object twice. Under
some circumstances (only seen on Windows), that could lead to program
crashes.
See :issue:`1587`.
- gevent can now be built using Cython 3.0a5 and newer. The PyPI
distribution uses this version.
The libev extension was incompatible with this. As part of this,
certain internal, undocumented names have been changed.
(Technically, gevent can be built with Cython 3.0a2 and above.
However, up through 3.0a4 compiling with Cython 3 results in
gevent's test for memory leaks failing. See `this Cython issue
<https://github.com/cython/cython/issues/3578>`_.)
See :issue:`1599`.
- Destroying a hub after joining it didn't necessarily clean up all
resources associated with the hub, especially if the hub had been
created in a secondary thread that was exiting. The hub and its parent
greenlet could be kept alive.
Now, destroying a hub drops the reference to the hub and ensures it
cannot be switched to again. (Though using a new blocking API call may
still create a new hub.)
Joining a hub also cleans up some (small) memory resources that might
have stuck around for longer before as well.
See :issue:`1601`.
- Fix some potential crashes under libuv when using
``gevent.signal_handler``. The crashes were seen running the test
suite and were non-deterministic.
See :issue:`1606`.
----